Verified, then logged.
Sun 22 Nov 2026 · 00:15 KST / 15:15 UTC한국어 | ENSearch /

Breaking01 / AI

HintErase targets subtitle removal despite detection errors

The model uses the original video and region hints to remove missed subtitles while preserving wrongly marked background.

A laptop on an office desk displays four bars of differing lengths, resembling probabilities for each choice, with a person seen from behind out of focus in the background. AI-generated image

HintErase researchers presented a video subtitle removal model on arXiv on 11 Nov, designed to work even when detection marks the wrong area.

According to the paper, the model receives the full, unmasked video alongside hints showing where to erase. During training, the researchers introduce hint errors resembling those made by real detectors. The target for removal follows the actual subtitle location.

The researchers said they trained the model to remove subtitles missing from the hints. It is also trained to preserve background that the hints wrongly cover. Text belonging to the original scene is meant to remain intact.

For long videos, the paper describes a processing pipeline that handles scene cuts and audio. It also handles missing hints and hints that shift between frames. The researchers designed its memory use to remain independent of video length.

The paper explains that existing methods rebuild the marked region and copy everything outside it. This carries over errors when detectors miss subtitle strokes or outlines. It also identifies the problem of marking background between letters for removal.

The abstract does not disclose numerical comparisons of removal performance.